Skip to content

agents-sdk.md

Latest commit

 

History

History
188 lines (129 loc) · 4.85 KB

agents-sdk.md

File metadata and controls

188 lines (129 loc) · 4.85 KB
Raw

Building with the SDK

Build MCP-powered agents with the Hugging Face agentic SDKs. The huggingface_hub (Python) and @huggingface/tiny-agents (JavaScript) libraries provide everything you need to connect LLMs to MCP tools.

Installation

pip install "huggingface_hub[mcp]"
npm install @huggingface/tiny-agents
# or
pnpm add @huggingface/tiny-agents

Quick Start: Run an Agent

The fastest way to get started is with the tiny-agents CLI:

tiny-agents run julien-c/flux-schnell-generator
npx @huggingface/tiny-agents run "julien-c/flux-schnell-generator"

This loads an agent from the tiny-agents collection, connects to its MCP servers, and starts an interactive chat.

Using the Agent Class

The Agent class manages the chat loop and MCP tool execution. It uses Inference Providers to run the LLM.

from huggingface_hub import Agent
import asyncio

agent = Agent(
    model="Qwen/Qwen2.5-72B-Instruct",
    provider="novita",
    servers=[
        {
            "type": "sse",
            "url": "https://evalstate-flux1-schnell.hf.space/gradio_api/mcp/sse"
        }
    ]
)

async def main():
    async for chunk in agent.run("Generate an image of a sunset"):
        if hasattr(chunk, 'choices'):
            delta = chunk.choices[0].delta
            if delta.content:
                print(delta.content, end="")

asyncio.run(main())

See the Agent reference for all options.

import { Agent } from "@huggingface/tiny-agents";

const agent = new Agent({
    model: "Qwen/Qwen2.5-72B-Instruct",
    provider: "novita",
    apiKey: process.env.HF_TOKEN,
    servers: [
        {
            type: "sse",
            url: "https://evalstate-flux1-schnell.hf.space/gradio_api/mcp/sse"
        }
    ]
});

await agent.loadTools();

for await (const chunk of agent.run("Generate an image of a sunset")) {
    if ("choices" in chunk) {
        const delta = chunk.choices[0]?.delta;
        if (delta.content) {
            console.log(delta.content);
        }
    }
}

See the tiny-agents documentation for all options.

Using MCPClient Directly

For more control, use MCPClient to manage MCP servers and tool calls directly.

import asyncio
from huggingface_hub import MCPClient

async def main():
    async with MCPClient(
        model="Qwen/Qwen2.5-72B-Instruct",
        provider="novita",
    ) as client:
        # Connect to an MCP server
        await client.add_mcp_server(
            type="sse", 
            url="https://evalstate-flux1-schnell.hf.space/gradio_api/mcp/sse"
        )
        
        # Process a request with tools
        messages = [{"role": "user", "content": "Generate an image of a sunset"}]
        
        async for chunk in client.process_single_turn_with_tools(messages):
            if hasattr(chunk, 'choices'):
                delta = chunk.choices[0].delta
                if delta.content:
                    print(delta.content, end="")

asyncio.run(main())

See the MCPClient reference for all options.

The JavaScript SDK uses the Agent class for MCP interactions. For lower-level control, see the @huggingface/mcp-client package.

Share Your Agent

Contribute agents to the tiny-agents collection on the Hub. Include:

  • agent.json - Agent configuration (required)
  • PROMPT.md or AGENTS.md - System prompt (optional)
  • EXAMPLES.md - Sample prompts and use cases (optional)

Learn More